TL;DR
OpenAI released its flagship model GPT-6 Astra on September 3, 2026 — its most intelligent and best-aligned model to date, and the first rated at the "Critical" level of cybersecurity capability under its Preparedness Framework. Standard API pricing is $10 per million input tokens and $50 per million output tokens, about 2.5x the previous GPT-5.6 Sol, with initial access limited to trusted institutions and vetted cybersecurity users.
Background
GPT-6 Astra was officially released by OpenAI on September 3, 2026 (early morning of September 4, Beijing time). In the preceding month, OpenAI spent considerable time explaining the model's safety profile: on August 7 it first disclosed that it could not rule out the model reaching the "Critical" cybersecurity capability threshold, and on September 1 it formally confirmed that Astra became OpenAI's first model rated Critical under the Preparedness Framework. At the launch, OpenAI president Greg Brockman described the model with the words "Welcome to the AGI era."
1. Training and core capabilities
Astra was pre-trained for the first time at the Stargate site in Texas using more than 100,000 GPUs, with everything from the data-center network to the inference kernels designed from scratch around that scale of training. It is also OpenAI's first model in which older-generation models participated deeply in the training process. OpenAI calls it "the world's best computer use model," shifting the launch narrative from "what the model can answer" to "how many tasks can be handed directly to AI."
Context and output: a 1.05-million-token context window, up to 128K tokens of output, and a knowledge cutoff of April 30, 2026. Codex gains a cross-window note mechanism that keeps notes and re-searches earlier messages and tool outputs when the context window fills up.
2. Key benchmark results
| Benchmark | Astra score | Comparison |
|---|---|---|
| ARC-AGI-3 (Provider Adapter setup) | 99.9% | 62.7% max under the Standard harness |
| FrontierMath Tier 4 | 97.6% | — |
| DeepSWE v1.1 | 74.1% | — |
| BenchCAD | 95.9% | — |
| Terminal-Bench Science | 64.6% | — |
| SRE-Bench reverse engineering (1 / up to 4 attempts) | 88.0% / 99.2% | GPT-5.6 Sol: 55.9% / 68.7% |
| MRCR v2 eight-needle retrieval (256K-512K / 512K-1M context) | 100% / 96.3% | Sol: 91.5% / 73.8% |
| ExploitBench | 100% | — |
Caveats: the 99.9% ARC-AGI-3 score is under OpenAI's own Provider Adapter setup (including two special settings, retained reasoning and compaction); the maximum under ARC Prize's standard harness is 62.7%. OpenAI says that on tasks both setups completed, the Provider Adapter ran about 3.66x faster and used 49% fewer tokens. FrontierMath was developed by Epoch AI, which says OpenAI funded its development and holds exclusive access to some of its problems.
3. Safety and alignment: the first Critical-grade model
Critical level means that, given suitable tools and access, Astra can autonomously discover previously unknown vulnerabilities and develop usable exploits in many hardened real-world systems without step-by-step human guidance, or design and execute an end-to-end attack strategy against a hardened target from a high-level objective alone.
OpenAI's internal safety tests: computer-use violation rate dropped from 22.0% for GPT-5.6 Sol to 2.4%; 0% in the test for bypassing restrictions to complete tasks; internal hallucination rate dropped from 9.4% to 2.0%. During testing, the model independently discovered and exploited two previously unknown zero-day vulnerabilities; OpenAI says it is disclosing them to the relevant maintainers.
To manage risk, OpenAI publicly acknowledged on August 18 that it had paused scaling up training (including two weeks of RL training planned for deployment); it deployed "misalignment monitoring" in production, observing the model's reasoning, actions, tool calls and full task trajectories, and can pause or terminate a task when it appears to deviate seriously from the user's original intent. OpenAI estimates the monitoring overhead alone amounts to about 20% of the compute of the inference being monitored.
4. Pricing and rollout
GPT-6 Astra standard API pricing: $10 per million input tokens, $1 cached input, $50 per million output tokens — about 2.5x GPT-5.6 Sol ($4 input, $20 output). Fast mode doubles the price ($20 input, $100 output) for up to roughly 2.5x execution speed. Long-context surcharge: for a single request above 272K input tokens, the full request is billed at 2x for input and 1.5x for output.
Rollout: initially open only to a small set of institutions in the Trusted Access/Daybreak program and vetted cybersecurity users; Plus, Pro, Business and Enterprise subscribers and the API receive access in the days following release; Pro, Business and Enterprise get the more capable Astra Pro (Enterprise defaults to off, requiring an admin to enable). The API model name is gpt-6-astra, also available on Amazon Bedrock.
5. The AGI debate
Brockman said "Welcome to the AGI era" at the launch and said his personal judgment is that "we have achieved" AGI; pressed on whether OpenAI officially declares AGI, he acknowledged AGI is more like a "mission concept" or "spiritual concept" than a precisely definable technical standard. CEO Sam Altman has said in interviews that AGI is "a very vaguely defined concept" and that he worries more about the next word: Superintelligence.
Common misconceptions and details
- "Released" does not mean "fully available." Astra was initially open only to trusted institutions; regular subscribers needed several days to gain access, and the capability boundaries differ by user.
- The 99.9% ARC-AGI-3 score depends on OpenAI's custom harness. Cross-model comparisons should use the ARC Prize standard harness figure (62.7%).
- Cybersecurity capabilities are tiered. The default production version still limits high-risk cyberattack requests, and advanced cybersecurity workflows roll out gradually only to vetted users in the Daybreak Blue program.
Summary
GPT-6 Astra is the flagship model OpenAI released on September 3, 2026 — the first rated Critical under the Preparedness Framework — priced 2.5x the previous generation with "computer use" as its core narrative. It marks a shift in frontier-model competition from "answering ability" to "autonomously completing tasks," while raising new issues of safety-monitoring cost (about 20% of inference compute) and tiered capability access.
Sources and references
- OpenAI official research page (GPT-6 Astra announcement): link
- 36Kr: GPT-6 Astra is live — has the AGI era really arrived?: link
- iFeng: GPT-6 Astra brings the 'AGI moment'? Industry and community still debate: link
- Convly AI: GPT-6 Astra Pricing, Benchmarks and What Actually Changed: link
- Wallstreetcn: Matching performance but raising price 2.5x: link